Rotor wing vibration adjustment prediction method and system based on deep learning

Abstract

The invention discloses a rotor vibration adjustment prediction method and system based on deep learning, and belongs to the technical field of rotor vibration control. The method comprises the following steps: receiving a vibration amplitude, a phase, a blade number and geometric parameters; analyzing the phase into sine and cosine components, interacting with a vibration value, converting geometric parameters into physical characteristics such as a span-chord ratio and rotational inertia, and constructing a 31-dimensional characteristic vector; a multi-layer perceptron model is trained for 2-8 blades, the input of the multi-layer perceptron model is the 31-dimensional feature vector, and the output of the multi-layer perceptron model is the balance weight adjustment amount of each non-reference blade; and performing feature processing on real-time data, inputting the processed data into a corresponding model, obtaining a prediction scheme, and estimating the vibration reduction amount. According to the method, high-precision, rapid and self-adaptive rotor vibration prediction and adjustment are realized by fusing a physical mechanism and deep learning, and the maintenance efficiency and the flight safety are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rotor vibration control, in particular to a rotor vibration adjustment prediction method and system based on deep learning. BACKGROUND

[0002] The rotor system is the core lifting and control component of an aircraft such as a helicopter or a drone, and the vibration level thereof is a key indicator for evaluating flight quality, safety and reliability. Excessive vibration not only affects the comfort of passengers and the service life of onboard equipment, but also can cause structural fatigue and serious safety accidents.

[0003] Currently, existing rotor vibration adjustment mainly relies on artificial experience or simplified mathematical models, such as the counterweight balancing method based on the influence coefficient matrix. The above methods have the following disadvantages: (1) low adjustment accuracy: the coupling effect of multiple blades (2-8 pieces) is complex, and traditional linear models cannot accurately predict the counterweight adjustment amount of each blade, resulting in unsatisfactory vibration suppression effect; (2) poor adaptability: the influence of blade geometric parameters (length, chord length, airfoil type and efficiency) on aerodynamic load and vibration phase is not systematically considered, and when the blade type is changed or the task load is modified, recalibration is required, which cannot dynamically adapt to different rotor configurations; (3) low efficiency: highly dependent on the experience of maintenance personnel, and each adjustment requires a cycle of "flight / test - measurement - calculation", which is time-consuming and dependent on expert experience, and cannot predict in real time during flight; (4) low degree of intelligence: unable to model multiple fault coupling scenarios such as "imbalance, misalignment, resonance", lacking risk assessment and visualization means. SUMMARY

[0004] The present application provides a rotor vibration adjustment prediction method and system based on deep learning, and a rotor aircraft configured with the system. The vibration adjustment prediction method can be applied to 2-8 blade configurations, taking into account geometric parameters and fault coupling, significantly improving the prediction accuracy of counterweight; achieving minute-level end-to-end prediction, reducing the number of test runs and human dependence; reducing the adjusted vibration to below 0.15 IPS, and providing risk level assessment.

[0005] Technical solution: The present application provides a rotor vibration adjustment prediction method based on deep learning, comprising the following steps: Step S1, parameter input: receiving the input vibration amplitude, vibration phase, number of blades n and geometric parameters of the aircraft rotor; the geometric parameters include blade length, chord length, area, mass, thickness and airfoil type, the airfoil type is selected from a symmetric airfoil database, the database contains thickness and aerodynamic efficiency data of widely used airfoils such as NACA0006 and NACA0012; the number of blades n of the aircraft rotor is 2-8 pieces; Step S2, feature engineering: the parameters input in step S1 are constructed into a 31-dimensional feature vector suitable for deep learning model processing; the 31-dimensional feature vector includes: 6-dimensional basic vibration features constructed based on vibration amplitude and vibration phase; 4-dimensional fault type features encoded by one-hot encoding of the current fault type; 7-dimensional blade number features encoded by one-hot encoding of the input 2-8 blade number; 8-dimensional blade geometry features converted from the input blade geometry parameters; 6-dimensional vibration-related features constructed from the perspectives of statistics, trends, and risk levels; Step S3, data generation: simulate the vibration response under different fault types based on the rotor dynamics principle to generate a training data set for training the deep learning model, simulate the vibration response of the rotor under various working conditions; the fault types include imbalance, misalignment, resonance, and normal state; Step S4, model training: for each type of 2-8 blade rotor of the aircraft, a corresponding deep learning model is constructed and trained, the deep learning model is a multi-layer perceptron model, the number of input layer nodes is consistent with the dimension of the 31-dimensional feature vector, the number of output layer nodes is equal to the number of blades minus one, used to predict the weight adjustment amount of the non-reference blade relative to the reference blade; the training goal of each model is to find the optimal weight adjustment amount required to reduce the vibration to the target value (such as <0.15 IPS) on each non-reference blade, which is obtained by inverse calculation in the deep learning model through the training data set data; Step S5, prediction output: input the 31-dimensional feature vector constructed in the feature engineering step into the trained deep learning model corresponding to the current number of blades, to obtain the weight adjustment amount of each non-reference blade, and estimate the vibration reduction amount after adjustment.

[0006] Further, the vibration phase is input in degrees and seconds format, which is automatically parsed into a decimal angle value in the feature engineering step, and then triangular decomposition is performed.

[0007] Further, the 31-dimensional feature vector is specifically composed as follows: Basic vibration features: including vibration amplitude, normalized phase, phase sine component, phase cosine component, interaction term of vibration amplitude and phase sine component, interaction term of vibration amplitude and phase cosine component, total of 6 dimensions; Fault type features: one-hot encoding is used to encode four states of imbalance, misalignment, resonance, and normal, total of 4 dimensions; Blade number features: one-hot encoding is used to encode the blade number n, total of 7 dimensions; Blade geometry features: including blade length, chord length, aspect ratio, area, mass, moment of inertia, airfoil efficiency, and thickness, total of 8 dimensions; Vibration-related features: including over-limit vibration sign, logarithmic transformation of vibration amplitude, power transformation of vibration amplitude, standard score of vibration amplitude relative to historical average, rate of change of vibration amplitude, and discrete level coding of vibration amplitude, a total of 6 dimensions, which enhance the characterization of vibration amplitude from multiple dimensions.

[0008] Further, the multi-layer perception model comprises an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer connected in turn; wherein: the input layer is a feature vector node of 31-dimensional features; the first hidden layer is 256 nodes, adopts ReLU function activation and connects Dropout layer; the second hidden layer is 128 nodes, adopts ReLU function activation and connects Dropout layer; the third hidden layer is 64 nodes, adopts ReLU function activation; the output layer is n-1 node, n is the current number of blades, adopts ReLU function activation to ensure that the output of the balance weight adjustment amount is a non-negative value.

[0009] Further, the model training step adopts Adam optimizer and mean square error loss function to optimize the model, and saves the model weight file corresponding to each number of blades and the parameters for feature standardization after the training is completed.

[0010] Further, the prediction output step further comprises a balance weight distribution sub-step: according to the vibration phase, the total balance weight predicted by the model is distributed to each non-reference blade according to the phase projection principle, and the weight is normalized, wherein the distribution weight is proportional to cos(the difference between the phase of each blade and the vibration phase)+1.

[0011] Further, it further comprises a risk assessment step: comparing the vibration value before prediction with a plurality of preset vibration threshold values, and outputting a "normal", "mild", "moderate", "severe" or "dangerous" level prompt.

[0012] The application also provides a rotor vibration adjustment prediction system based on deep learning, comprising: A parameter input module is used to receive the vibration amplitude, vibration phase, number of 2-8 blades and geometric parameters of each blade of the aircraft rotor; the parameter input module comprises a degree-minute-second parser for automatically converting the vibration phase input by the user in degree-minute-second format into decimal angle; An input verification module is used to check whether the input vibration amplitude is within a preset range, such as 0.00-5.00 IPS, and whether the degree-minute-second components of the vibration phase are within a legal range.

[0013] A feature engineering module is used to generate a 31-dimensional feature vector suitable for deep learning model processing from the input parameters; A data generation module is used to generate training data under multiple fault types based on the principle of rotor dynamics; A model storage and calling module is configured to store the deep learning model and standardized parameters of the deep learning model trained for n blades respectively, wherein the deep learning model is a multi-layer perceptron, the input layer has 31 nodes, and the output layer has n-1 nodes, wherein n is the number of blades. A prediction output module is configured to input the real-time generated feature vector into the corresponding trained deep learning model to obtain the counterweight adjustment amount of each non-reference blade and estimate the vibration reduction amount after adjustment. A graphical user interface module is configured to provide a parameter input interface and display the prediction results and risk assessment information.

[0014] Further, the prediction output module further includes a counterweight distribution sub-module configured to distribute the total counterweight amount to each non-reference blade according to the vibration phase by the cosine projection method and output a specific gram-level counterweight scheme.

[0015] Further, the standardized parameters saved by the model storage and calling module are used to perform standardization processing on the real-time generated feature vector before prediction to match the data distribution during model training.

[0016] Advantages: Compared with the prior art, the present application has the following technical effects and advantages: (1) High prediction accuracy: by deeply integrating the physical characteristics of the rotor (geometric parameter conversion, phase decomposition) and the nonlinear mapping ability of deep learning, the vibration response under complex working conditions can be accurately described, the counterweight prediction accuracy is significantly improved compared with traditional methods, and the vibration can be adjusted to below 0.15 IPS to meet the airworthiness requirements.

[0017] (2) High adjustment efficiency: end-to-end intelligent prediction is realized on a minute level, the user only needs to input basic parameters to obtain the counterweight scheme, which greatly reduces the test flight times and maintenance time, and reduces the dependence on expert experience.

[0018] (3) Strong adaptability: the geometric characteristics of the blades are converted into features recognizable by the model through feature engineering, so that the same model framework can adapt to rotor systems with different numbers of blades (2-8) and different geometric parameters, without the need to recalibrate for each configuration, and the generalization ability is good.

[0019] (4) High degree of intelligence: integrated fault type coding, risk assessment and visual output, providing comprehensive decision support for maintenance personnel, improving the scientificity and safety of maintenance work. BRIEF DESCRIPTION OF DRAWINGS

[0020] Fig. 1 A flowchart of the rotor vibration adjustment prediction method based on deep learning of the present application; Fig. 2 A neural network structure diagram of the multi-layer perceptron model of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings and examples.

[0022] Example 1 Method flow Reference Figs. 1-2 , a rotor vibration adjustment prediction method based on deep learning, the specific steps are as follows: Step S1, parameter input: collect the rotor vibration amplitude of the aircraft 0.42 IPS, the vibration phase is 45°30′15″, the number of blades is 4, and the geometric parameters of each blade are: length 1.8 m, chord length 0.18 m, and airfoil NACA0012 (airfoil efficiency 0.88, airfoil thickness 0.12 m); Step S2, feature engineering: the core of this step is to convert the original parameters into a 31-dimensional feature vector rich in information and suitable for deep learning model processing. The specific construction is as follows: Basic vibration features (6 dimensions): based on vibration amplitude V (0.42 IPS) and vibration phase Φ (45°30′15″). First, analyze the vibration phase to decimal 45.504°, and then normalize: φ = 45.504 / 360 ≈ 0.1264. Calculate the phase sine component sin(2πφ) ≈ 0.713 and the cosine component cos(2πφ) ≈ 0.701. Finally, this group of features includes: ① vibration amplitude 0.42; ② normalized phase 0.1264; ③ phase sine component 0.713; ④ phase cosine component 0.701; ⑤ vibration amplitude and phase sine component interaction item 0.42 * 0.713 ≈ 0.300; ⑥ vibration amplitude and phase cosine component interaction item 0.42 * 0.701 ≈ 0.296. This 6-dimensional feature collectively describes the intensity and direction information of the vibration vector.

[0023] Fault type features (4 dimensions): use one-hot encoding to identify the current fault state (unbalance, misalignment, resonance, normal). Assuming that the current diagnosis is an “unbalance” fault, it is represented as a four-dimensional vector [1, 0, 0, 0] using one-hot encoding.

[0024] Blade number features (7 dimensions): use one-hot encoding to identify the number of blades of the rotor (2 to 8 blades), and for a configuration of 4 blades, use one-hot encoding to represent [0, 0, 1, 0, 0, 0, 0] (corresponding to 2, 3, 4, 5, 6, 7, 8 blades).

[0025] Blade geometry (8 dimensions): Including blade length, chord length, aspect ratio, area, mass, moment of inertia, airfoil efficiency, and thickness, the input blade geometry parameters are converted into physical characteristics: 1. Length 1.8 m; 2. Chord length 0.18 m; 3. Aspect ratio 10.0; 4. Area 0.324 m²; 5. Mass 1.5 kg; 6. Estimated moment of inertia 0.972 kg·m²; 7. Airfoil efficiency 0.88; 8. Airfoil thickness 0.12 m.

[0026] Vibration-related features (6 dimensions): This group of features deepens the understanding of the vibration state from the perspectives of statistics, trends, and risk levels, which are calculated as follows: Exceedance of safety threshold: Binary feature (0 or 1) indicating whether the current vibration value exceeds the pre-set safety / alarm threshold. The primary safety threshold is set to 0.30 IPS, so this feature is 1 because V = 0.42 > 0.30.

[0027] Log transformation: log 10 (vibration amplitude + 0.01), this transformation compresses the scale of the data, making the model's sensitivity to high-amplitude vibrations more moderate, focusing more on order-of-magnitude changes, and ensuring that the transformation is valid when the vibration value is 0. log 10 (0.42 + 0.01) ≈ log 10 (0.43) ≈ -0.367.

[0028] Power transformation: vibration value^1.5. This transformation amplifies the impact of high-amplitude vibrations, making the model more concerned about large vibrations that need urgent treatment, complementing the log transformation. 0.42^1.5 ≈ 0.273.

[0029] Standard score: (current value - historical mean) / historical standard deviation. This feature quantifies the deviation of the current vibration from the "normal baseline" of the rotor, effectively identifying abnormal mutations. Assuming the historical healthy vibration mean of this rotor μ_hist = 0.10 IPS and the standard deviation σ_hist = 0.05 IPS, the standard score is (0.42 - 0.10) / 0.05 = 6.4.

[0030] Change rate: The difference between this vibration amplitude and the last measurement value. This feature captures the dynamic change trend of the vibration, which is an important indicator for judging whether the fault is rapidly deteriorating. Assuming the previous measurement value is 0.38 IPS, the change rate is 0.42 - 0.38 = 0.04.

[0031] Discrete level coding: According to multiple thresholds (0, 0.15, 0.3, 1, 2.5, 5), the vibration value is coded into discrete levels (such as 0, 1, 2, 3, 4), which comprehensively and intuitively reflects the severity of vibration. The vibration value 0.42 IPS meets 0.30≤ V<1.00, so it is coded as level 2 (corresponding to “moderate”).

[0032] Feature combination: The above-mentioned five groups of 31-dimensional features are standardized and combined into a unified feature vector, which is input into the deep learning model.

[0033] Step S3, data generation: First, forward simulation: based on the blade momentum theory and the multi-body dynamics equation, a high-fidelity rotor system digital twin simulation model is constructed, different fault conditions (such as specific mass distribution of each blade to simulate imbalance, hub installation angle deviation to simulate misalignment, structural stiffness and damping parameter change to simulate resonance trend, etc.) are set, and the corresponding blade geometric parameters are input. Run the simulation to calculate the steady-state vibration response of the rotor system under the fault state, i.e. the simulation vibration amplitude and phase; Second, reverse optimization calculation: the simulation vibration value is suppressed to below the preset target safety threshold (for example, <0.15 IPS) as the optimization goal, and virtual counterweights are applied to the non-reference blades for adjustment in the simulation environment. An optimization algorithm (such as gradient descent or exhaustive search) is used to solve a set of optimal counterweight adjustment combinations [ΔM_2, ΔM_3,..., ΔM_n] that can make the final simulation vibration the lowest (and meet the target) through iterative calculation (where ΔM_i represents the mass adjustment of the i-th non-reference blade relative to the reference blade, in grams); Finally, sample construction: each complete simulation working condition generates a training sample pair. The fault state, blade geometric parameters, and simulation measured vibration amplitude and phase are used as the original input, which is processed into a 31-dimensional feature vector through the feature engineering process described in step S2, as the input feature of the sample. The optimal counterweight adjustment combination calculated by the physical model is used as the true label of the sample. By generating a large number of samples covering “imbalance”, “misalignment”, “resonance”, “normal” states and working conditions, a “feature-label” sample pair dataset for model training is constructed.

[0034] Step S4, model training: using the above-mentioned large-scale "feature-label" sample pairs, the internal weights and bias parameters of the multilayer perceptron model are continuously adjusted through the back propagation algorithm and gradient descent optimization (such as using the Adam optimizer). The optimization goal is to minimize the difference between the model's predicted weight adjustment and the optimal weight adjustment calculated by the physical model (measured by mean square error MSE). Through this process, the model learns the high-dimensional nonlinear mapping relationship between complex fault features and accurate physical adjustment, thereby obtaining the prediction ability. To adapt to the configuration of 2 to 8 different number of blades, an independent modeling strategy is adopted. For each number of blades n, the training data generated under the corresponding configuration is used to independently complete the above data preparation and model training process, and a dedicated n-blade multilayer perceptron model is obtained. All trained model weight files and their corresponding feature standardization parameters (mean and standard deviation) are persistently stored in the system's model library. The model structure is: input layer (31 nodes, i.e., 31-dimensional feature vector) → first hidden layer (256 nodes, ReLU function activation) → Dropout layer (dropout rate 20%) → second hidden layer (128 nodes, ReLU function activation) → Dropout layer (dropout rate 20%) → third hidden layer (64 nodes, ReLU function activation) → Dropout layer (dropout rate 10%) → fourth hidden layer (32 nodes, ReLU function activation) → Dropout (dropout rate 10%) → fifth hidden layer (n-1 nodes) → output layer (non-negative weight vector, node number depends on task setting). Using the training data generated in step S3, the mean square error is used as the loss function, and the Adam is used as the optimizer for training until the model converges on the validation set. Save the trained model weights and feature standardization parameters. Wherein, n is the number of blades of the current rotor (for example, 4 blades output 3 values), and the ReLU function activation ensures that the weight adjustment predicted by the model is non-negative, which conforms to the physical reality (only increase the weight).

[0035] Step S5, prediction output: input the real-time 31-dimensional feature vector generated in step S2 into the trained 4-blade multilayer perceptron model. Dynamically load and call the corresponding dedicated model and standardization parameters to achieve adaptive and high-precision prediction for different rotor configurations. The model outputs a 3-dimensional vector, for example, [2.1, 0.8, 1.5], with the unit defaulting to grams (g). This represents the weight that needs to be added to the 2nd, 3rd, and 4th non-reference blades (assuming the 1st blade as the reference). The system estimates that this adjustment can reduce the vibration by about 0.34 IPS, and the expected residual vibration is 0.08 IPS.

[0036] Step S6, counterweight distribution: according to the vibration phase (45.504°), the total counterweight is finely distributed by using the phase projection principle. The difference between the blade azimuth angle and the vibration phase is calculated, and the final counterweight scheme is obtained by calculating and normalizing the distribution weight proportional to [cos(phase difference) + 1], such as: 2# blade + 2.1g @ XX°, 3# blade + 0.8g @ YY°, 4# blade + 1.5g @ ZZ°.

[0037] Step S7, risk assessment: the system compares the initial vibration value 0.42 IPS with the built-in threshold (for example: below 0.15 IPS is "normal", 0.15-0.30 is "slight", 0.30-1.00 is "moderate", 1.00-2.50 is "severe", and above 2.50 is "dangerous"), and gives a "moderate" risk prompt. At the same time, according to the predicted residual vibration 0.08 IPS, a green prompt of "adjusted to reach the safety standard" is displayed.

[0038] Embodiment 2 system structure The rotor vibration adjustment prediction system comprises: A parameter input module connected with the airborne vibration sensor and the maintenance panel, which receives vibration values, phases, blade numbers and geometric parameters; A feature engineering module integrating a degree-second resolver and an airfoil database, which outputs a 31-dimensional feature vector; A data generation module deployed on the ground station, which generates training samples through a high-fidelity rotor dynamics model; A model storage and calling module which stores model weight files and standardization parameters corresponding to 2-8 blades; A prediction output module comprising a counterweight distribution submodule, which outputs a gram-level counterweight scheme; A graphical user interface module which provides an input box, a level lamp, a one-key prediction button, and displays prediction results and risk assessment information; A communication interface which supports ARINC-429, CAN bus or 4G / 5G link, and sends the counterweight scheme to the counterweight execution mechanism.

[0039] As can be seen from the above embodiments, the present application realizes high-precision, fast and self-adaptive vibration adjustment through data-driven method without changing the structure of the aircraft rotor, solves the problems of low precision, poor adaptability and low efficiency of traditional methods, and has significant industrial practical value.

Claims

1. A deep learning-based method for predicting rotor vibration adjustment, characterized in that, Includes the following steps: Step S1, parameter input: Receive the vibration amplitude, vibration phase, number of blades n, and geometric parameters of each blade of the aircraft rotor, where 2≤n≤8; Step S2, Feature Engineering: Construct the parameters input in Step S1 into a 31-dimensional feature vector suitable for deep learning model processing; the 31-dimensional feature vector includes: a 6-dimensional basic vibration feature based on vibration amplitude and vibration phase; a 4-dimensional fault type feature encoded by one-hot encoding of the current fault type; a 7-dimensional blade number feature represented by the input blade number n using one-hot encoding; an 8-dimensional blade geometric feature transformed from the input blade geometric parameters; and a 6-dimensional vibration-related feature constructed from statistical, trend, and risk level perspectives. Step S3, Data Generation: Based on the principle of rotor dynamics, simulate the vibration response under different fault types to generate a training dataset for training the deep learning model; the fault types include imbalance, misalignment, resonance, and normal state; Step S4, Model Training: For each type of aircraft rotor blade in the n blade types, a corresponding deep learning model is constructed and trained. The deep learning model is a multilayer perceptron model, whose input layer node number is consistent with the dimension of the 31-dimensional feature vector, and whose output layer node number is equal to the number of blades minus 1. It is used to predict the weight adjustment amount of non-reference blades relative to the reference blade. Step S5, Prediction Output: Input the 31-dimensional feature vector constructed by the feature engineering step into the trained deep learning model corresponding to the current number of blades to obtain the counterweight adjustment amount of each non-reference blade and estimate the vibration reduction amount after adjustment.

2. The method according to claim 1, characterized in that, The vibration phase is input in degrees, minutes, and seconds format. In the feature engineering step, it is automatically parsed into a decimal angle value, and then trigonometrically decomposed into sine and cosine components.

3. The method according to claim 2, characterized in that, The 31-dimensional feature vector is specifically constructed as follows: Basic vibration characteristics: including vibration amplitude, normalized phase, phase sinusoidal component, phase cosine component, interaction term between vibration amplitude and phase sinusoidal component, and interaction term between vibration amplitude and phase cosine component, totaling 6 dimensions; Fault type characteristics: Four states—imbalance, misalignment, resonance, and normal—are encoded using one-hot coding, resulting in a total of 4 dimensions; Blade quantity characteristics: The number of n blades is encoded using one-hot encoding, with a total of 7 dimensions; Blade geometry features include blade length, chord length, aspect ratio, area, mass, moment of inertia, airfoil efficiency, and thickness, totaling 8 dimensions. Vibration-related features include: over-limit vibration indicators, logarithmic transformation of vibration amplitude, power transformation of vibration amplitude, standard fraction of vibration amplitude relative to historical average, rate of change of vibration amplitude, and discrete level coding of vibration amplitude, totaling 6 dimensions.

4. The method according to claim 1, characterized in that, The multilayer perceptron model comprises an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer connected in sequence; wherein: the input layer consists of feature vector nodes with 31-dimensional features; the first hidden layer consists of 256 nodes, activated by the ReLU function and connected to a Dropout layer; the second hidden layer consists of 128 nodes, activated by the ReLU function and connected to a Dropout layer; the third hidden layer consists of 64 nodes, activated by the ReLU function; and the output layer consists of n-1 nodes, where n is the current number of blades, activated by the ReLU function to ensure that the output weight adjustment is a non-negative value.

5. The method according to claim 1, characterized in that, The model training step uses the Adam optimizer and mean squared error loss function to optimize the model, and after training, saves the model weight file corresponding to the number of blades and the parameters used for feature standardization.

6. The method according to claim 1, characterized in that, The prediction output step further includes a counterweight allocation sub-step: based on the vibration phase, the total counterweight predicted by the model is allocated to each non-reference blade according to the phase projection principle, and the weights are normalized, wherein the allocation weight is proportional to the cosine value of each blade phase minus the vibration phase plus 1.

7. The method according to claim 1, characterized in that, It also includes a risk assessment step: comparing the vibration value before prediction with multiple preset vibration thresholds and outputting a "normal", "slight", "moderate", "severe" or "dangerous" level prompt.

8. A deep learning-based rotor vibration adjustment prediction system, characterized in that, include: The parameter input module is used to receive the vibration amplitude, vibration phase, number of blades n, and geometric parameters of each blade of the aircraft rotor; the parameter input module includes a degree-minute-second resolver, which is used to automatically convert the vibration phase in degree-minute-second format input by the user into a decimal angle; The input verification module is used to verify whether the input vibration amplitude is within the preset range and whether the vibration phase is within the legal range. The data generation module is used to generate training data for various fault types based on the principles of rotor dynamics. The feature engineering module is used to generate 31-dimensional feature vectors suitable for deep learning model processing from the input parameters; The model storage and retrieval module is used to store the deep learning models trained for each of the n propeller blades and their standardized parameters. The deep learning model is a multilayer perceptron model with 31 nodes in the input layer and n-1 nodes in the output layer. The prediction output module is used to input the real-time generated feature vectors into the corresponding trained deep learning model to obtain the counterweight adjustment amount of each non-reference blade and estimate the vibration reduction amount after adjustment. The graphical user interface module provides a parameter input interface and displays prediction results and risk assessment information.

9. The system according to claim 8, characterized in that, The prediction output module also includes a counterweight allocation submodule, which is used to allocate the total counterweight to each non-reference blade according to the vibration phase using the cosine projection method, and output a specific gram-level counterweight scheme.

10. The system according to claim 8 or 9, characterized in that, The standardized parameters saved by the model storage and retrieval module are used to standardize the real-time generated feature vectors before prediction in order to match the data distribution during model training.

Images